Curated Open-Source AI Repositories

“Best-in-class data connectors for PDFs, Notion, SQL, and enterprise data with hybrid retrieval.”

“My long-term memory primitive — the palace model with wings and drawers maps neatly to tenant-isolated memory in our sovereign stack. 58K stars because it treats memory as architecture. When NOT to use: Not a vector DB — pair with qdrant/pgvector for retrieval, use palace for structured cross-session recall.”

“My RAG ingest front-door — Firecrawl extracts clean markdown where LlamaParse misses tables. Pairs with LlamaIndex Workflows for event-driven retrieval.”

“Graphify holds 0.3% hallucination on our GST cross-check where pgvector alone drifted to 1.2% — the graph is the guardrail.”

“Claude Code → Codex handoffs break for me; ai-memory fixes it with one binary, no vector DB, SQLite + Markdown versioned in git. v1.32.2 trending Aug 27 at 5,055 stars validates the file-not-vector approach for Gujarat SMEs on 4G. Rating: 4.6/5. When NOT to use: Avoid if you need semantic chunk retrieval at scale — this is file-level handoff memory, not pgvector GraphRAG.”

“The 92.5k-star fix for “every session starts from scratch” — claude-mem captures everything, compresses with AI and re-injects per agent. Works with Claude Code/OpenClaw/Codex/Gemini/Hermes/Copilot/OpenCode, which is why it leads AI memory trending at +218 today. Rating: 4.8/5. When NOT to use: Not if you already use file-based handoff (ai-memory) — claude-mem is session-capture heavy, heavier than git Markdown.”

“I split memory (MemPalace), RAG (pgvector HNSW 42ms) and skills (catalog) across three systems — OpenViking unifies all three into one self-evolving store. Its memory+RAG+skills convergence is the architecture I am migrating my Junagadh ledger toward for Gujarat SME agents.”

“I roll out shared agent setups for 5-10 staff Gujarat SME teams where everyone hand-rolls prompts and context dies per seat. TencentDB extracting reusable Skills plus CodeGraph from past sessions is the cold-start fix I prescribe from Junagadh, and I test its proxy mode on the Rs 6K VPS with zero-code client switching.”

“Reliable citations for India compliance (DPDP/GST) — grounds Hindi/Gujarati docs before answering, vs hallucinated RAG.”

“Blazing fast vector lookup with rich metadata payload filtering, written in Rust with minimal memory footprint.”

“Fixes stateless chat — remembers Junagadh client history across sessions, essential for vernacular voice + WhatsApp agents.”

“Zero-setup embedded database that lets you spin up local vector search in 3 lines of Python.”

“Most RAG is vector-only and forgets why an answer exists. Semantica adds graph-native provenance so an audit can trace context lineage — that is why it held #1 on Aug 11 before agency-agents reclaimed it. Rating: 4.3/5. When NOT to use: Skip for tiny single-doc Q&A — graph provenance pays off at team scale and compliance audits, not at “answer this PDF”.”

“Code search that actually understands imports, call graphs and cross-file edges — Tree-sitter + Memgraph beats chunk-embedding alone for refactoring. The +7 rank rebound #13→#6 on Aug 12 tracks real dev pain. Rating: 4.5/5. When NOT to use: Overkill for single-file scripts — gains show at multi-language monorepo scale where structure matters.”

“For SME RAG I need provenance + eval — OpenKB bundles ingestion, vector store and retrieval scoring openly. #1 Trending AI Memory Aug 29 at 4.0k (+930 today) because teams are done with black-box RAG. Rating: 4.3/5. When NOT to use: Skip for single-PDF chat — OpenKB pays off as a shared org KB, not ad-hoc demo.”

“I keep 90-day JSONL per tenant — ECC is the smallest repo that gets agent memory + eval right together. Its performance harness is what I benchmark my 500-sample replay against.”